Reuse quality of open polymer datasets: a reproducible cross-repository audit (2020–2025)

Polymer scientists have been quietly flooding the open-data ecosystem. Between 2020 and 2025, more than 53,000 openly deposited polymer datasets accumulated across the eight major generalist research-data repositories that serve every scientific discipline — Zenodo, Figshare, Mendeley Data, Harvard Dataverse, OSF, 4TU.ResearchData, Dryad and Fordatis. Now, for the first time, someone has audited them all. The verdict is startling in its simplicity: what determines whether a polymer dataset can actually be reused is not the science inside it, nor the sub-field it belongs to, nor even the year it was deposited — it is which repository the researcher happened to choose. Where you deposit, it turns out, matters roughly ten times more than anything else about your data.

The audit matters because polymer informatics and machine-learning models are increasingly trained on aggregations of openly deposited datasets, and the value of such an aggregation is bounded by whether a downstream researcher can find, understand and re-execute each contributing dataset. That is precisely what the FAIR principles — Findability, Accessibility, Interoperability, Reusability — formalised in 2016, and what funder mandates and journal data-availability policies now demand. Yet polymer researchers have had nowhere domain-specific to go. Polymer Genome, PI1M and OpenPoly are curated, read-only platforms that accept no community deposition, while the domain-deep crystallographic infrastructures NOMAD and the Materials Project carry less than 0.1 percent polymer-relevant content. The experimental polymer community has therefore been directed, de facto, to the same generalist repositories used across every other discipline — and until now, nobody had measured what happens to polymer data once it lands there.

The study, published as a reproducible cross-repository audit, harvested all eight repositories with a controlled vocabulary of 74 polymer- and sustainability-related keywords, assembling a corpus of 53,038 records spanning 2020 to 2025. Each record was scored on 11 binary FAIR-aligned criteria: presence of a DOI, an open license, a sufficiently rich abstract, keywords, a classifiable material system, a method description, listed variables, identifiable units, at least one file in an open format, a README, and a code link tracked separately. The score is deliberately thin — a corpus-level screening instrument rather than a per-record FAIR diagnosis — and the entire pipeline is reproducible byte-identically from the deposited raw harvest: a clean end-to-end run regenerated 32 of 33 published outputs identically in roughly six minutes on commodity hardware, with the one exception being a live URL check that is reproducible in procedure but not in bytes.

The first headline finding is a structural discontinuity in metadata completeness. Bibliographic fields — the ones sufficient for citation — are near saturation: a DOI is present in 93.2 percent of records, a rich abstract in 87.4 percent, keywords in 86.1 percent, and any license in 82.9 percent. But the fields that determine whether a downstream researcher can actually reuse the data collapse below 50 percent: a measurement-method description appears in just 41.3 percent of records, a synthesis-method description in 32.2 percent, listed variables in 21.5 percent, at least one open-format file in 10.3 percent, identifiable units in 8.3 percent, a README in 4.3 percent and a code link in a mere 1.2 percent. The discontinuity around the 50 percent mark separates two functional groups of fields — and it is the bottom group that carries the reuse value.

The second headline finding is that repository identity alone explains 37.9 percent of reuse-score variance — a large effect by conventional statistical thresholds, and an order of magnitude greater than the variance attributable to publication year, sub-domain or sustainability framing. The Kruskal-Wallis omnibus test returns H = 20,104 with p below 10 to the power of minus 300, and all 28 pairwise contrasts between repositories are highly significant after Bonferroni correction. A four-tier structure emerges: Dryad, Zenodo and Figshare occupy the high tier with median scores of five or above; 4TU.ResearchData, Fordatis and Mendeley Data sit in the middle; Harvard Dataverse falls to a median of three; and OSF collapses to a median of one. The largest contrast is OSF versus Dryad, with Cliff’s delta of minus 0.977 — a randomly drawn Dryad record outranks a randomly drawn OSF record in 98.2 percent of pairings. Notably, Mendeley Data, despite contributing 42 percent of the corpus by volume, sits only in the middle tier.

The third headline finding is a decoupling of volume from quality. Annual deposit volume rose monotonically from 5,905 records in 2020 to 14,374 in 2025 — a 143 percent increase, accelerating in 2024 and 2025 with the largest year-on-year jump of 40 percent. Over the same window, the mean reuse score rose from 3.84 to 4.14 — a relative gain of just 7.7 percent, less than one-third of one score point on the 0-to-10 range. The record-level Spearman correlation between publication year and score is statistically significant on more than 52,000 records, but the effect size is trivially small. The open polymer-data ecosystem, in other words, is widening without deepening: repositories are accepting substantially more deposits each year, but the average documentation quality of those deposits is essentially flat, and the roughly nineteen-fold gap between volume growth and quality growth means the signal-to-noise ratio of the ecosystem is not improving in line with its size.

Perhaps the most counter-intuitive single result concerns non-isocyanate polyurethanes, or NIPUs — the green-chemistry alternative to conventional polyurethanes, which rely on phosgene-derived diisocyanates classified as REACH SVHC substances. NIPU chemistry is mature, with vibrant research programmes on polyhydroxyurethanes produced by polyaddition of bis-cyclic carbonates and polyamines, often from bio-based feedstocks such as vegetable oils and terpenes. Yet NIPU-keyword-matched records — 8,647 of them, 16.3 percent of the corpus — carry a median reuse score of 2, lower than the median of 4 for conventional polyurethane records, with a medium effect size. The paradox dissolves under a leave-one-out sensitivity test: when the two lowest-tier repositories, OSF and Dataverse, are excluded, the disadvantage collapses to a non-significant negligible effect. Seventy-eight percent of NIPU records sit in OSF and Dataverse, while 86 percent of conventional polyurethane records sit in Figshare and Zenodo. The gap is platform-mediated, not behavioural — a finding about where the NIPU community deposits, not about how it documents.

Two further results sharpen the picture. Sustainable and bio-based polymer datasets — bio-based, biodegradable and recycled or circular polymers — are modestly better documented than conventional ones, with a small but consistent documentation premium that is never negative across all eight repositories. Yet they represent only 6.8 percent of the corpus, a lower bound given the keyword-recall asymmetry, despite the enormous publication volume in these areas driven by European Green Deal policy, corporate ESG mandates and microplastic concern. Sustainable-polymer chemistry, the audit suggests, is published far more than it is deposited as FAIR-conformant data. Separately, the audit identified and resolved a structural OSF file-level FAIR gap: 53 percent of OSF records are files returned by the search API with empty creators, license and DOI — an architectural consequence of OSF’s parent-attachment data model that pre-dates FAIR. A three-pass parent-inheritance workaround, using only the relationship link already present in the API response, lifted file-record creator coverage from 0 percent to 96.8 percent with no additional calls, leaving a 3.2 percent structural ceiling.

Within the score itself, the audit resisted the temptation to single out one high-leverage field. The raw criterion-total correlation assigns method description the largest coefficient, but this is inflated by a part-whole dependency — the total score already contains the criterion being correlated. Under the corrected item-total correlation, no criterion exceeds 0.41 and method description ranks sixth of eleven. What survives the correction is not a single dominant field but a structural divide: the near-saturated bibliographic basics offer little headroom, while the cluster of reuse-enabling content and file-level criteria — method, variables, units, open files, README, code — is sparsely populated. The binding constraint on aggregate reuse quality is the breadth of under-populated content metadata, not any one criterion.

The implications converge on the operator side. A repository-agnostic intervention is unlikely to recover the variance attributable to repository identity, while operator-level interventions — broadening and enforcing the cluster of reuse-enabling content fields at deposit time, and exposing file-level metadata that the API model already supports — can address it directly. The Generalist Repository Ecosystem Initiative, which publishes a shared DataCite 4.6 metadata baseline adopted by seven of the eight audited repositories, appears necessary but not sufficient: GREI-participating repositories span the full quality range from Dryad’s mean of 6.98 to OSF’s 1.53, indicating that schema agreement without enforcement leaves outcomes essentially unchanged. The practical advice for polymer researchers is concrete: choose Figshare, Zenodo, 4TU.ResearchData or Mendeley Data over OSF or Dataverse for primary deposition of experimental datasets. And the methodology itself — the deposited corpus, code and audit suite, re-verifiable in about six minutes — is designed as a calibration benchmark that other domains, from life sciences to energy materials, can re-apply to ask the same question of their own open-data ecosystems.

Subject of Research: Reuse quality of open polymer datasets: a reproducible cross-repository audit (2020–2025)

Article Title: Reuse quality of open polymer datasets: a reproducible cross-repository audit (2020–2025)

Article References: Meyer, W. (2026). Reuse quality of open polymer datasets: a reproducible cross-repository audit (2020–2025). Machine Learning with Applications, 26, Article 101044. https://doi.org/10.1016/j.mlwa.2026.101044

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101044

Keywords: Reuse, quality, open, polymer, datasets, reproducible, cross-repository, audit, scientific research